Is @Albertsons delivery service the worst or is it just me? I get countless deliveries and I rarely have any issues delivering to my Apartment except theirs. The drivers they use are either not shown the delivery directions or are not following them. Only one of my last half dozen deliveries was without issues.
1
2
163
Julias Shaw retweeted
Gradient descent for SKILL.md files sounds interesting, maybe a bit complex but it's becoming a real part of agent harness.
SkillOpt is one of the first papers to treat markdown skill files as trainable parameters and provides a proper optimization framework for them.
A few things I learned that you should consider too.
1. The validation gate is the only thing that matters in a self-editing loop.
Held-out set, strict improvement, ties rejected. End-to-end, their best skills land with 1 to 4 accepted edits total. If your "self-improving agent" is accepting most of what it proposes, you're shipping slop.
2. Bounded edits are better than full rewrites. 4 to 8 edits per step is the sweet spot.
Remove the budget and performance collapses. This is the textual analog of learning rate, and it transfers to any LLM-as-author loop. If you're using an agent to refactor your docs, your prompts, or your skills, cap the diff size.
3. Compactness wins. Median final skill: ~920 tokens.
Skills do not need to be long. They need to be high-signal. Most skill files I see are bloated because length feels like effort. It isn't.
4. The harness is becoming less important; the skill is becoming more important.
A Codex-trained skill ported into Claude Code hit +59.7 points on SpreadsheetBench. Procedural knowledge is more general than the runtime that
produced it.
5. Frozen model + trained context is the practical adaptation.
GPT-5.4-nano with a SkillOpt'd skill ≈ frontier behavior on procedural benchmarks. Cheaper, portable, inspectable, zero inference-time cost. This is
the answer to "how do we adapt a frontier model for our domain" for almost everyone who isn't training their own models.
6. Verification is the bottleneck.
Every gate in this paper depends on an auto-grader. That works for benchmarks. It fails for writing, design, and strategy, exactly the open-ended work we want to automate. Whoever builds the verifier for open-ended tasks owns the next stage.
There are also two leassons I learned while shipping v2.3.0 of my Context Engineering Agent Skills repo, measured across composer-2, claude-opus-4-7,
gpt-5.5, and gemini-3.1-pro via the @cursor_ai SDK:
- Description and body are two different surfaces. The router only sees the description. The agent sees the body once activated. They can quietly disagree, and only end-to-end task tests catch it.
- Aggregate accuracy is the wrong unit. When I rewrote three descriptions, the corpus average moved ~1pp. Individual skills moved 23–25pp. Per-skill effect size is where the action is.
Also, in Feb 2026 I shared a piece called Personal Brain OS arguing that the markdown file is a first-class substrate for agent state. SkillOpt is the optimizer-shaped version of that same argument: not "store memory in files" but "treat files as trainable parameters with proper optimization machinery around them." That's the move from static to measured.
The fast/slow split they describe already lives implicitly in the digital-brain-skill repo:
- voice-guide and tone-of-voice.md are slow-state (rarely touched)
- posts.jsonl and bookmarks.jsonl are fast-state
What SkillOpt adds that I didn't have is a protected section invariant, a structural guarantee that fast edits cannot overwrite slow lessons. Removing that mechanism cost them 22 points on SpreadsheetBench. Worth borrowing.
If you're building agents, SkillOpt: Executive Strategy for Self-Evolving Agent Skills is a good paper to read: arxiv.org/pdf/2605.23904
I just sponsored @simonw. Go sponsor your open source dependencies! github.com/sponsors/simonw?s…
2
54
Watch The AI Doc trailer and join the movement. youtube.com/watch?v=xkPbV3IR…
1
49
Julias Shaw retweeted
Videos recorded, edited, and ready to roll.
Build an AI Powered Instagram Clone, The Video Course, Launching January 7th!
#MyElixirStatus
I’ve been a Netflix customer since they were only shipping DVDs but between rising subscription prices, other quality streaming services, and them not planning to support the Vision Pro I think it’s time to cancel my account. #AppleVisionPro
193
If you’re wondering what prescriptions are supported by the Zeiss inserts for Apple Vision Pro here’s the details. Most non-prism lenses within -11.00d to +6.75D #AppleVisionPro
1
222
And now for the waiting... Due to technical difficulties my order took a while to get in but I'm expecting delivery by Feb 16th! #AppleVisionPro
157
"Goals are for people who care about winning once. Systems are for people who care about winning repeatedly."
-@JamesClear
62
Elixir & Phoenix LiveView are awesome! Here's a great example from @JohnElmLabs showing how to build an Instagram clone in 20 minutes with ZERO JavaScript... youtube.com/watch?v=4cnmyQJT…
and so it begins...
P.S. The Apple Vision Pro SDK is now available!
#AppleVisionPro
147
If I owned movie theaters I would be terrified about what Vision Pro and the devices to follow it mean.
For some perspective, $3,500 is less than the yearly price of 2 Grande Mochas at Starbucks a day.
84
Julias Shaw retweeted
I spent 10% of my life contributing to the development of the #VisionPro while I worked at Apple as a Neurotechnology Prototyping Researcher in the Technology Development Group. It’s the longest I’ve ever worked on a single effort. I’m proud and relieved that it’s finally announced. I’ve been working on AR and VR for ten years, and in many ways, this is a culmination of the whole industry into a single product. I’m thankful I helped make it real, and I’m open to consulting and taking calls if you’re looking to enter the space or refine your strategy.
The work I did supported the foundational development of Vision Pro, the mindfulness experiences, ▇▇▇▇▇▇ products, and also more ambitious moonshot research with neurotechnology. Like, predicting you’ll click on something before you do, basically mind reading. I was there for 3.5 years and left at the end of 2021, so I’m excited to experience how the last two years brought everything together. I’m really curious what made the cut and what will be released later on.
Specifically, I’m proud of contributing to the initial vision, strategy and direction of the ▇▇▇▇▇▇ program for Vision Pro. The work I did on a small team helped green light that product category, and I think it could have significant global impact one day.
The large majority of work I did at Apple is under NDA, and was spread across a wide range of topics and approaches. But a few things have become public through patents which I can cite and paraphrase below.
Generally as a whole, a lot of the work I did involved detecting the mental state of users based on data from their body and brain when they were in immersive experiences.
So, a user is in a mixed reality or virtual reality experience, and AI models are trying to predict if you are feeling curious, mind wandering, scared, paying attention, remembering a past experience, or some other cognitive state. And these may be inferred through measurements like eye tracking, electrical activity in the brain, heart beats and rhythms, muscle activity, blood density in the brain, blood pressure, skin conductance etc.
There were a lot of tricks involved to make specific predictions possible, which the handful of patents I’m named on go into detail about. One of the coolest results involved predicting a user was going to click on something before they actually did. That was a ton of work and something I’m proud of. Your pupil reacts before you click in part because you expect something will happen after you click. So you can create biofeedback with a user's brain by monitoring their eye behavior, and redesigning the UI in real time to create more of this anticipatory pupil response. It’s a crude brain computer interface via the eyes, but very cool. And I’d take that over invasive brain surgery any day.
Other tricks to infer cognitive state involved quickly flashing visuals or sounds to a user in ways they may not perceive, and then measuring their reaction to it.
Another patent goes into details about using machine learning and signals from the body and brain to predict how focused, or relaxed you are, or how well you are learning. And then updating virtual environments to enhance those states. So, imagine an adaptive immersive environment that helps you learn, or work, or relax by changing what you’re seeing and hearing in the background.
All of these details are publicly available in patents, and were carefully written to not leak anything. There was a ton of other stuff I was involved with, and hopefully more of it will see the light of day eventually.
A lot of people have waited a long time for this product. But it’s still one step forward on the road to VR. And it’s going to take until the end of this decade for the industry to fully catch up to the grand vision for this tech.
Again, I’m open to consulting work and taking calls if your business is looking to enter the space or refine your strategy. Mostly, I’m proud and relieved this has finally been announced. It’s been over five years since I started working on this, and I spent a significant portion of my life on it, as did an army of other designers and engineers. I hope the whole is greater than the sum of the parts and Vision Pro blows your mind.
It would be incredible if @OpenBCI created a custom EEG head strap for the Apple Vision Pro! #VisionPro
89
Yes! A laptop with AR glasses instead of a screen! pcworld.com/article/1919392
1
38











